1. Enhance & Explore: CH-1015 Lausanne adel.aziz@epfl.ch http://people.epfl.ch/adel.aziz Adel Aziz joint work with Julien Herzen, Ruben Merz, Seva Shneer, and Patrick Thiran September 21 st 2011 LCA an Adaptive Algorithm to Maximize the Utility of Wireless Networks

2.  Objective  Maximize given utility function  Challenges vs. related work  Work on existing MAC  No network-wide message passing  Wireless capacity is unknown a priori 1/14 Problem Statement (Max-Weight based [1] ) (IFA [2] ) (GAP [3] ) [1] Proutière et al., CISS, 2008 [2] Gambiroza et al., MobiCom, 2004 [3] Mancuso et al., Infocom, 2010

3. WLAN Setting  Inter-flow problem  Optimally allocate resources Multi-Hop Setting  Intra-flow problem  Avoid congestion 2/14 Motivation [3] Heusse et al., Infocom, 2003 [3]

4.  Flow =  No scaling problem  Architecture of node j  Information to set at node j at time slot n  Rate allocation vector:  Information to monitor at GW at time slot n  Measured throughput: 3/14 Network Abstraction

5.  Useful information at time slot n  Measured throughput:  Capacity region: (unknown a priori)  Rate allocation vector:  Last stable allocation:  Utility function:  Level set: 4/14 Network Abstraction x1x1 X*X* x2x2

6.  Starts from IEEE 802.11 allocation oThroughput vector: x[0] = ρ[0] {rate allocation} oLevel set of utility μ[0]: L(x[0], μ[0]) oRemember allocation:r[0] = x[0]  Enhance phase: (Find the next targeted level set) oIf (x[n-1] = ρ[n-1] )  μ[n] = full-size gradient ascent oElse  μ[n] = halved-size gradient ascent  Explore phase: (Find the next allocation) oPick point ρ[n] randomly oIf (x[n] = ρ[n] )  Remember new allocation: r[n] = x[n]  Go to Enhance phase oElse  Repeat Explore phase at most N times, then move to Enhance μ[0] μ[2] μ[3] μ [1] Truncated Gaussian pdf μ [4] Example: N = 2 ‘Enhance & Explore’ in WLAN 5/14

7.  Theorem for E&E  Utility of never decreases through time  converges to an allocation of maximal utility  Converges for any initial rate allocation  Assumptions for the proof  Fixed capacity region  Coordinate-convex capacity region  Much weaker than convexity!  Future work  Study speed of convergence 6/14 Optimality Theorem

8.  GW  E&E decides the per-flow rate allocation  One-hop nodes  E&E rate-limits one-hop nodes  Multi-hop nodes  EZ-flow [1] rate-limits multi-hop nodes 7/14 Complete Solution: E&E + EZ [1] Aziz et al., CoNEXT 2009

9.  Based on Click [1] with MultiflowDispatcher [2]  Creation of 5 new elements  MFQueue  MFLeakyBucket  EEscheduler  EEadapter  EZFlow  Evaluation with  Asus routers  Ns-3 [2] Schiöberg et al., SyClick, 2009 8/14 Practical Implementation [1] Kohler et al., Transactions on Computer Systems, 2000

10.  Deployment map:  Without E&E:  With E&E: (proportional fairness) 9/14 Experimental Results in WLAN

11.  Inter-flow fairness  Appropriate queuing is needed (Fair Queuing [1] )  … but it is not enough 10/14 [1] Demers et al., Sigcomm’89 Starvation in Mesh Networks

12.  Deployment map:  Without E&E:  With E&E: (proportional fairness) 11/14 Practical Results in Mesh Networks 1-hop flow 3-hop flow 1-hop flow 3-hop flow

13.  Ns-3 simulator  Re-use of same Click elements  More controlled environment  Possible estimation of capacity  Computation of optimum 12/14 Simulation Results in WLAN

14.  Include downstream traffic  Study and improve the speed of convergence  Analyze new distributions for the Explore phase  Study the interaction with rate adaptation 13/14 Future Work

15.  Wireless networks suffer from  Intra-flow problem (e.g., congestion)  Inter-flow problem (e.g., unfairness)  Time-variability  Difficulty/impossibility characterizing the capacity region  Need adaptive algorithms to maximize a desired utility  E&E solves the inter-flow problem in WLAN  Combining E&E and EZ-Flow in mesh networks  Solves both the inter-flow and intra-flow problem  Avoids network-wide message passing  Does not modify networking stack 14/14 Conclusion